Abstract: We present Generalizable NeRF Transformer (GNT), a pure, unified transformer-based architecture that efficiently reconstructs Neural Radiance Fields (NeRFs) on the fly from source views. Unlike prior works on NeRF that optimize a per-scene implicit representation by inverting a handcrafted rendering equation, GNT achieves generalizable neural scene representation and rendering, by encapsulating two transformers-based stages. The first stage of GNT, called view transformer, leverages multi-view geometry as an inductive bias for attention-based scene representation, and predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views. The second stage of GNT, named ray transformer, renders novel views by ray marching and directly decodes the sequence of sampled point features using the attention mechanism. Our experiments demonstrate that when optimized on a single scene, GNT can successfully reconstruct NeRF without explicit rendering formula, and even improve the PSNR by ~1.3 dB↑ on complex scenes due to the learnable ray renderer. When trained across various scenes, GNT consistently achieves the state-of-the-art performance when transferring to forward-facing LLFF dataset (LPIPS ~20%↓, SSIM ~25%↑) and synthetic blender dataset (LPIPS ~20%↓, SSIM ~4%↑). In addition, we show that depth and occlusion can be inferred from the learned attention maps, which implies that the pure attention mechanism is capable of learning a physically-grounded rendering process. All these results bring us one step closer to the tantalizing hope of utilizing transformers as the "universal modeling tool" even for graphics.

Abstract: Accelerating communication for users with severe motor and speech impairments, in particular for eye-gaze Augmentative and Alternative Communication (AAC) device users, is a long-standing area of research. However, observation of such users' communication over extended durations has been limited. This case study presents the real-world experience of developing and field-testing a tool for observing and curating the gaze typing-based communication of a consented eye-gaze AAC user with amyotrophic lateral sclerosis (ALS) from the perspective of researchers at the intersection of HCI and artificial intelligence (AI). With the intent to observe and accelerate eye-gaze typed communication, we designed a tool and a protocol called the SpeakFaster Observer to measure everyday conversational text entry by the consenting gaze-typing user, as well as several consenting conversation partners of the AAC user. We detail the design of the Observer software and data curation protocol, along with considerations for privacy protection. The deployment of the data protocol from November 2021 to April 2022 yielded a rich dataset of gaze-based AAC text entry in everyday context, consisting of 130+ hours of gaze keypresses and 5.5k+ curated speech utterances from the AAC user and the conversation partners. We present the key statistics of the data, including the speed (8.1±3.9 words per minute) and keypress saving rate (-0.18±0.87) of gaze typing, patterns of of utterance repetition and reuse, as well as the temporal dynamics of conversation turn-taking in gaze-based communication. We share our findings and also open source our data collections tools for furthering research in this domain.

Abstract: Improving performance of deep networks in data limited regimes has warranted much attention. In this work, we empirically show that “winning tickets” (small subnetworks) obtained via magnitude pruning based on the lottery ticket hypothesis, apart from being sparse are also effective recognizers in data limited regimes. Based on extensive experiments, we find that in low data regimes (datasets of 50-100 examples per class), sparse winning tickets substantially outperform the original dense networks. This approach, when combined with augmentations or fine-tuning from a self-supervised backbone network, shows further improvements in performance by as much as 16% (absolute) on low sample datasets and longtailed classification. Further, sparse winning tickets are more robust to synthetic noise and distribution shifts compared to their dense counterparts. Our analysis of winning tickets on small datasets indicates that, though sparse, the networks retain density in the initial layers and their representations are more generalizable.

Abstract: Motivated by the need for accelerating text entry in augmentative and alternative communication (AAC) for people with severe motor impairments, we propose a paradigm in which phrases are abbreviated aggressively as primarily word-initial letters. Our approach is to expand the abbreviations into full-phrase options by leveraging conversation context with the power of pretrained large language models (LLMs). Through zero-shot, few-shot, and fine-tuning experiments on four public conversation datasets, we show that for replies to the initial turn of a dialog, an LLM with 64B parameters is able to accurately expand over 70% of phrases with abbreviation length up to 10, leading to an effective keystroke saving rate of up to 77% on these expansions. Including a small amount of context in the form of a single conversation turn more than doubles abbreviation expansion accuracies compared to having no context, an effect that is more pronounced for longer phrases. Additionally, the robustness of the models against typo noise can be enhanced through fine-tuning on noisy data.

Abstract: Amyotrophic Lateral Sclerosis (ALS) disease severity is usually measured using the subjective, questionnaire-based revised ALS Functional Rating Scale (ALSFRS-R). Objective measures of disease severity would be powerful tools for evaluating real-world drug effectiveness, efficacy in clinical trials, and for identifying participants for cohort studies. We developed a machine learning (ML) based objective measure for ALS disease severity based on voice samples and accelerometer measurements from a four-year longitudinal dataset. 584 people living with ALS consented and carried out prescribed speaking and limb-based tasks. 542 participants contributed 5814 voice recordings, and 350 contributed 13,009 accelerometer samples, while simultaneously measuring ALSFRS-R scores. Using these data, we trained ML models to predict bulbar-related and limb-related ALSFRS-R scores. On the test set (n = 109 participants) the voice models achieved a multiclass AUC of 0.86 (95% CI, 0.85–0.88) on speech ALSFRS-R prediction, whereas the accelerometer models achieved a median multiclass AUC of 0.73 on 6 limb-related functions. The correlations across functions observed in self-reported ALSFRS-R scores were preserved in ML-derived scores. We used these models and self-reported ALSFRS-R scores to evaluate the real-world effects of edaravone, a drug approved for use in ALS. In the cohort of 54 test participants who received edaravone as part of their usual care, the ML-derived scores were consistent with the self-reported ALSFRS-R scores. At the individual level, the continuous ML-derived score can capture gradual changes that are absent in the integer ALSFRS-R scores. This demonstrates the value of these tools for assessing disease severity and, potentially, drug effects.

Abstract: Drug discovery for diseases such as Parkinson’s disease are impeded by the lack of screenable cellular phenotypes. We present an unbiased phenotypic profiling platform that combines automated cell culture, high-content imaging, Cell Painting, and deep learning. We applied this platform to primary fibroblasts from 91 Parkinson’s disease patients and matched healthy controls, creating the largest publicly available Cell Painting image dataset to date at 48 terabytes. We use fixed weights from a convolutional deep neural network trained on ImageNet to generate deep embeddings from each image and train machine learning models to detect morphological disease phenotypes. Our platform’s robustness and sensitivity allow the detection of individual-specific variation with high fidelity across batches and plate layouts. Lastly, our models confidently separate LRRK2 and sporadic Parkinson’s disease lines from healthy controls (receiver operating characteristic area under curve 0.79 (0.08 standard deviation)), supporting the capacity of this platform for complex disease modeling and drug screening applications.

Abstract: Recent advances in self-supervision have dramatically improved the quality of speech representations. However, deployment of state-of-the-art embedding models on devices has been restricted due to their limited public availability and large resource footprint. Our work addresses these issues by publicly releasing a collection of paralinguistic speech models that are small and near state-of-the-art performance. Our approach is based on knowledge distillation, and our models are distilled on public data only. We explore different architectures and thoroughly evaluate our models on the Non-Semantic Speech (NOSS) benchmark. Our largest distilled model is less than 15% the size of the original model (314MB vs 2.2GB), achieves over 96% the accuracy on 6 of 7 tasks, and is trained on 6.5% the data. The smallest model is 1% in size (22MB) and achieves over 90% the accuracy on 6 of 7 tasks. Our models outperform the open source Wav2Vec 2.0 model on 6 of 7 tasks, and our smallest model outperforms the open source Wav2Vec 2.0 on both emotion recognition tasks despite being 7% the size.

Abstract: Automatic classification of disordered speech can provide an objective tool for identifying the presence and severity of speech impairment. Classification approaches can also help identify hard-to-recognize speech samples to teach ASR systems about the variable manifestations of impaired speech. Here, we develop and compare different deep learning techniques to classify the intelligibility of disordered speech on selected phrases. We collected samples from a diverse set of 661 speakers with a variety of self-reported disorders speaking 29 words or phrases, which were rated by speech-language pathologists for their overall intelligibility using a five-point Likert scale. We then evaluated classifiers developed using 3 approaches: (1) a convolutional neural network (CNN) trained for the task, (2) classifiers trained on non-semantic speech representations from CNNs that used an unsupervised objective [1], and (3) classifiers trained on the acoustic (encoder) embeddings from an ASR system trained on typical speech [2]. We found that the ASR encoder's embeddings considerably outperform the other two on detecting and classifying disordered speech. Further analysis shows that the ASR embeddings cluster speech by the spoken phrase, while the non-semantic embeddings cluster speech by speaker. Also, longer phrases are more indicative of intelligibility deficits than single words.

Abstract: Integrated Gradients (IG) is a commonly used feature attribution method for deep neural networks. While IG has many desirable properties, the method often produces spurious/noisy pixel attributions in regions that are not related to the predicted class when applied to visual models. While this has been previously noted, most existing solutions are aimed at addressing the symptoms by explicitly reducing the noise in the resulting attributions. In this work, we show that one of the causes of the problem is the accumulation of noise along the IG path. To minimize the effect of this source of noise, we propose adapting the attribution path itself -- conditioning the path not just on the image but also on the model being explained. We introduce Adaptive Path Methods (APMs) as a generalization of path methods, and Guided IG as a specific instance of an APM. Empirically, Guided IG creates saliency maps better aligned with the model's prediction and the input image that is being explained. We show through qualitative and quantitative experiments that Guided IG outperforms other, related methods in nearly every experiment.

Abstract: Symbolic techniques based on Satisfiability Modulo Theory (SMT) solvers have been proposed for analyzing and verifying neural network properties, but their usage has been fairly limited owing to their poor scalability with larger networks. In this work, we propose a technique for combining gradient-based methods with symbolic techniques to scale such analyses and demonstrate its application for model explanation. In particular, we apply this technique to identify minimal regions in an input that are most relevant for a neural network's prediction. Our approach uses gradient information (based on Integrated Gradients) to focus on a subset of neurons in the first layer, which allows our technique to scale to large networks. The corresponding SMT constraints encode the minimal input mask discovery problem such that after masking the input, the activations of the selected neurons are still above a threshold. After solving for the minimal masks, our approach scores the mask regions to generate a relative ordering of the features within the mask. This produces a saliency map which explains" where a model is looking" when making a prediction. We evaluate our technique on three datasets-MNIST, ImageNet, and Beer Reviews, and demonstrate both quantitatively and qualitatively that the regions generated by our approach are sparser and achieve higher saliency scores compared to the gradient-based methods alone.

Abstract: Diabetic retinopathy (DR) screening is instrumental in preventing blindness, but faces a scaling challenge as the number of diabetic patients rises. Risk stratification for the development of DR may help optimize screening intervals to reduce costs while improving vision-related outcomes. We created and validated two versions of a deep learning system (DLS) to predict the development of mild-or-worse ("Mild+") DR in diabetic patients undergoing DR screening. The two versions used either three-fields or a single field of color fundus photographs (CFPs) as input. The training set was derived from 575,431 eyes, of which 28,899 had known 2-year outcome, and the remaining were used to augment the training process via multi-task learning. Validation was performed on both an internal validation set (set A; 7,976 eyes; 3,678 with known outcome) and an external validation set (set B; 4,762 eyes; 2,345 with known outcome). For predicting 2-year development of DR, the 3-field DLS had an area under the receiver operating characteristic curve (AUC) of 0.79 (95%CI, 0.78-0.81) on validation set A. On validation set B (which contained only a single field), the 1-field DLS's AUC was 0.70 (95%CI, 0.67-0.74). The DLS was prognostic even after adjusting for available risk factors (p < 0.001). When added to the risk factors, the 3-field DLS improved the AUC from 0.72 (95%CI, 0.68-0.76) to 0.81 (95%CI, 0.77-0.84) in validation set A, and the 1-field DLS improved the AUC from 0.62 (95%CI, 0.58-0.66) to 0.71 (95%CI, 0.68-0.75) in validation set B. The DLSs in this study identified prognostic information for DR development from CFPs. This information is independent of and more informative than the available risk factors.

Abstract: Model explanation techniques play a critical role in understanding the source of a model's performance and making its decisions transparent. Here we investigate if explanation techniques can also be used as a mechanism for scientific discovery. We make three contributions: first, we propose a framework to convert predictions from explanation techniques to a mechanism of discovery. Second, we show how generative models in combination with black-box predictors can be used to generate hypotheses (without human priors) that can be critically examined. Third, with these techniques we study classification models for retinal images predicting Diabetic Macular Edema (DME), where recent work showed that a CNN trained on these images is likely learning novel features in the image. We demonstrate that the proposed framework is able to explain the underlying scientific mechanism, thus bridging the gap between the model's performance and human understanding.

Abstract: We study the attribution problem for deep networks applied to perception tasks. For vision tasks, attribution techniques attribute the prediction of a network to the pixels of the input image. We propose a new technique called Blur Integrated Gradients (Blur IG). This technique has several advantages over other methods. First, it can tell at what scale a network recognizes an object. It produces scores in the scale/frequency dimension, that we find captures interesting phenomena. Second, it satisfies the scale-space axioms, which imply that it employs perturbations that are free of artifact. We therefore produce explanations that are cleaner and consistent with the operation of deep networks. Third, it eliminates the need for baseline parameter for Integrated Gradients for perception tasks. This is desirable because the choice of baseline has a significant effect on the explanations. We compare the proposed technique against previous techniques and demonstrate application on three tasks: ImageNet object recognition, Diabetic Retinopathy prediction, and AudioSet audio event identification. Code and examples are at https://github.com/PAIR-code/saliency.

Abstract: Center-involved diabetic macular edema (ci-DME) is a major cause of vision loss. Although the gold standard for diagnosis involves 3D imaging, 2D imaging by fundus photography is usually used in screening settings, resulting in high false-positive and false-negative calls. To address this, we train a deep learning model to predict ci-DME from fundus photographs, with an ROC–AUC of 0.89 (95% CI: 0.87–0.91), corresponding to 85% sensitivity at 80% specificity. In comparison, retinal specialists have similar sensitivities (82–85%), but only half the specificity (45–50%, p < 0.001). Our model can also detect the presence of intraretinal fluid (AUC: 0.81; 95% CI: 0.81–0.86) and subretinal fluid (AUC 0.88; 95% CI: 0.85–0.91). Using deep learning to make predictions via simple 2D images without sophisticated 3D-imaging equipment and with better than specialist performance, has broad relevance to many other applications in medical imaging.

Abstract: Owing to the invasiveness of diagnostic tests for anaemia and the costs associated with screening for it, the condition is often undetected. Here, we show that anaemia can be detected via machine-learning algorithms trained using retinal fundus images, study participant metadata (including race or ethnicity, age, sex and blood pressure) or the combination of both data types (images and study participant metadata). In a validation dataset of 11,388 study participants from the UK Biobank, the metadata-only, fundus-image-only and combined models predicted haemoglobin concentration (in g dl–1) with mean absolute error values of 0.73 (95% confidence interval: 0.72–0.74), 0.67 (0.66–0.68) and 0.63 (0.62–0.64), respectively, and with areas under the receiver operating characteristic curve (AUC) values of 0.74 (0.71–0.76), 0.87 (0.85–0.89) and 0.88 (0.86–0.89), respectively. For 539 study participants with self-reported diabetes, the combined model predicted haemoglobin concentration with a mean absolute error of 0.73 (0.68–0.78) and anaemia an AUC of 0.89 (0.85–0.93). Automated anaemia screening on the basis of fundus images could particularly aid patients with diabetes undergoing regular retinal imaging and for whom anaemia can increase morbidity and mortality risks.

Abstract: Advances in automation and imaging have made it possible to capture large image datasets for experiments that span multiple weeks with multiple experimental batches of data. However, accurate biological comparisons across the batches is challenged by the batch-to-batch variation due to uncontrollable experimental noise (e.g., different stain intensity or illumination conditions). To mediate the batch variation (i.e. the batch effect), we developed a batch equalization method that can transfer images from one batch to another while preserving the biological phenotype. The equalization method is trained as a generative adversarial network (GAN), using the StarGAN architecture that has shown considerable ability in doing style transfer for consumer images. After incorporating an additional objective that disentangles batch effect from biological features using an existing GAN framework, we show that the equalized images have less batch information as determined by a batch-prediction task and perform better in a biologically relevant task (e.g., Mechanism of Action prediction).

Abstract: Confounding variables are a well known source of nuisance in biomedical studies. They present an even greater challenge when we combine them with black-box machine learning techniques that operate on raw data. This work presents two case studies. In one, we discovered biases arising from systematic errors in the data generation process. In the other, we found a spurious source of signal unrelated to the prediction task at hand. In both cases, our prediction models performed well but under careful examination hidden confounders and biases were revealed. These are cautionary tales on the limits of using machine learning techniques on raw data from scientific experiments.

Abstract: The etiological underpinnings of many CNS disorders are not well understood. This is likely due to the fact that individual diseases aggregate numerous pathological subtypes, each associated with a complex landscape of genetic risk factors. To overcome these challenges, researchers are integrating novel data types from numerous patients, including imaging studies capturing broadly applicable features from patient-derived materials. These datasets, when combined with machine learning, potentially hold the power to elucidate the subtle patterns that stratify patients by shared pathology. In this study, we interrogated whether high-content imaging of primary skin fibroblasts, using the Cell Painting method, could reveal disease-relevant information among patients. First, we showed that technical features such as batch/plate type, plate, and location within a plate lead to detectable nuisance signals, as revealed by a pre-trained deep neural network and analysis with deep image embeddings. Using a plate design and image acquisition strategy that accounts for these variables, we performed a pilot study with 12 healthy controls and 12 subjects affected by the severe genetic neurological disorder spinal muscular atrophy (SMA), and evaluated whether a convolutional neural network (CNN) generated using a subset of the cells could distinguish disease states on cells from the remaining unseen control–SMA pair. Our results indicate that these two populations could effectively be differentiated from one another and that model selectivity is insensitive to batch/plate type. One caveat is that the samples were also largely separated by source. These findings lay a foundation for how to conduct future studies exploring diseases with more complex genetic contributions and unknown subtypes.

Abstract: Each year, the treatment decisions for more than 230,000 breast cancer patients in the U.S. hinge on whether the cancer has metastasized away from the breast. Metastasis detection is currently performed by pathologists reviewing large expanses of biological tissues. This process is labor intensive and error-prone. We present a framework to automatically detect and localize tumors as small as 100 x 100 pixels in gigapixel microscopy images sized 100,000 x 100,000 pixels. Our method leverages a convolutional neural network (CNN) architecture and obtains state-of-the-art results on the Camelyon16 dataset in the challenging lesion-level tumor detection task. At 8 false positives per image, we detect 92.4% of the tumors, relative to 82.7% by the previous best automated approach. For comparison, a human pathologist attempting exhaustive search achieved 73.2% sensitivity. We achieve image-level AUC scores above 97% on both the Camelyon16 test set and an independent set of 110 slides. In addition, we discover that two slides in the Camelyon16 training set were erroneously labeled normal. Our approach could considerably reduce false negative rates in metastasis detection.

Abstract: Question How does the performance of an automated deep learning algorithm compare with manual grading by ophthalmologists for identifying diabetic retinopathy in retinal fundus photographs?
Finding In 2 validation sets of 9963 images and 1748 images, at the operating point selected for high specificity, the algorithm had 90.3% and 87.0% sensitivity and 98.1% and 98.5% specificity for detecting referable diabetic retinopathy, defined as moderate or worse diabetic retinopathy or referable macular edema by the majority decision of a panel of at least 7 US board-certified ophthalmologists. At the operating point selected for high sensitivity, the algorithm had 97.5% and 96.1% sensitivity and 93.4% and 93.9% specificity in the 2 validation sets.
Meaning Deep learning algorithms had high sensitivity and specificity for detecting diabetic retinopathy and macular edema in retinal fundus photographs.

Abstract: Recent captioning models are limited in their ability to scale and describe concepts unseen in paired image-text corpora. We propose the Novel Object Captioner (NOC), a deep visual semantic captioning model that can describe a large number of object categories not present in existing image-caption datasets. Our model takes advantage of external sources -- labeled images from object recognition datasets, and semantic knowledge extracted from unannotated text. We propose minimizing a joint objective which can learn from these diverse data sources and leverage distributional semantic embeddings, enabling the model to generalize and describe novel objects outside of image-caption datasets. We demonstrate that our model exploits semantic information to generate captions for hundreds of object categories in the ImageNet object recognition dataset that are not observed in MSCOCO image-caption training data, as well as many categories that are observed very rarely. Both automatic evaluations and human judgements show that our model considerably outperforms prior work in being able to describe many more categories of objects.

Abstract: Long videos captured by consumers are typically tied to some of the most important moments of their lives, yet ironically are often the least frequently watched. The time required to initially retrieve and watch sections can be daunting. In this work we propose novel techniques for summarizing and annotating long videos. Existing video summarization techniques focus exclusively on identifying keyframes and subshots, however evaluating these summarized videos is a challenging task. Our work proposes methods to generate visual summaries of long videos, and in addition proposes techniques to annotate and generate textual summaries of the videos using recurrent networks. Interesting segments of long video are extracted based on image quality as well as cinematographic and consumer preference. Key frames from the most impactful segments are converted to textual annotations using sequential encoding and decoding deep learning models. Our summarization technique is benchmarked on the VideoSet dataset, and evaluated by humans for informative and linguistic content. We believe this to be the first fully automatic method capable of simultaneous visual and textual summarization of long consumer videos.

Abstract: This paper investigates how linguistic knowledge mined from large text corpora can aid the generation of natural language descriptions of videos. Specifically, we integrate both a neural language model and distributional semantics trained on large text corpora into a recent LSTM-based architecture for video description. We evaluate our approach on a collection of Youtube videos as well as two large movie description datasets showing significant improvements in grammaticality while modestly improving descriptive quality.

Abstract: While recent deep neural network models have achieved promising results on the image captioning task, they rely largely on the availability of corpora with paired image and sentence captions to describe objects in context. In this work, we propose the Deep Compositional Captioner (DCC) to address the task of generating descriptions of novel objects which are not present in paired image-sentence datasets. Our method achieves this by leveraging large object recognition datasets and external text corpora and by transferring knowledge between semantically similar concepts. Current deep caption models can only describe objects contained in paired image-sentence corpora, despite the fact that they are pre-trained with large object recognition datasets, namely ImageNet. In contrast, our model can compose sentences that describe novel objects and their interactions with other objects. We demonstrate our model's ability to describe novel concepts by empirically evaluating its performance on MSCOCO and show qualitative results on ImageNet images of objects for which no paired image-caption data exist. Further, we extend our approach to generate descriptions of objects in video clips. Our results show that DCC has distinct advantages over existing image and video captioning approaches for generating descriptions of new objects in context.

Abstract: Real-world videos often have complex dynamics; and methods for generating open-domain video descriptions should be sensitive to temporal structure and allow both input (sequence of frames) and output (sequence of words) of variable length. To approach this problem, we propose a novel end-to-end sequence-to-sequence model to generate captions for videos. For this we exploit recurrent neural networks, specifically LSTMs, which have demonstrated state-of-the-art performance in image caption generation. Our LSTM model is trained on video-sentence pairs and learns to associate a sequence of video frames to a sequence of words in order to generate a description of the event in the video clip. Our model naturally is able to learn the temporal structure of the sequence of frames as well as the sequence model of the generated sentences, i.e. a language model. We evaluate several variants of our model that exploit different visual features on a standard set of YouTube videos and two movie description datasets (M-VAD and MPII-MD).

Abstract: Solving the visual symbol grounding problem has long been a goal of artificial intelligence. The field appears to be advancing closer to this goal with recent breakthroughs in deep learning for natural language grounding in static images. In this paper, we propose to translate videos directly to sentences using a unified deep neural network with both convolutional and recurrent structure. Described video datasets are scarce, and most existing methods have been applied to toy domains with a small vocabulary of possible words. By transferring knowledge from 1.2M+ images with category labels and 100,000+ images with captions, our method is able to create sentence descriptions of open-domain videos with large vocabularies. We compare our approach with recent work using language generation metrics, subject, verb, and object prediction accuracy, and a human evaluation.

Abstract: Models based on deep convolutional networks have dominated recent image interpretation tasks; we investigate whether models which are also recurrent, or "temporally deep", are effective for tasks involving sequences, visual and otherwise. We develop a novel recurrent convolutional architecture suitable for large-scale visual learning which is end-to-end trainable, and demonstrate the value of these models on benchmark video recognition tasks, image description and retrieval problems, and video narration challenges. In contrast to current models which assume a fixed spatio-temporal receptive field or simple temporal averaging for sequential processing, recurrent convolutional models are "doubly deep"' in that they can be compositional in spatial and temporal "layers". Such models may have advantages when target concepts are complex and/or training data are limited. Learning long-term dependencies is possible when nonlinearities are incorporated into the network state updates. Long-term RNN models are appealing in that they directly can map variable-length inputs (e.g., video frames) to variable length outputs (e.g., natural language text) and can model complex temporal dynamics; yet they can be optimized with backpropagation. Our recurrent long-term models are directly connected to modern visual convnet models and can be jointly trained to simultaneously learn temporal dynamics and convolutional perceptual representations. Our results show such models have distinct advantages over state-of-the-art models for recognition or generation which are separately defined and/or optimized.

Abstract: This paper integrates techniques in natural language processing and computer vision to improve recognition and description of entities and activities in real-world videos. We propose a strategy for generating textual descriptions of videos by using a factor graph to combine visual detections with language statistics. We use state-of-the-art visual recognition systems to obtain confidences on entities, activities, and scenes present in the video. Our factor graph model combines these detection confidences with probabilistic knowledge mined from text corpora to estimate the most likely subject, verb, object, and place. Results on YouTube videos show that our approach improves both the joint detection of these latent, diverse sentence components and the detection of some individual components when compared to using the vision system alone, as well as over a previous n-gram language-modeling approach. The joint detection allows us to automatically generate more accurate, richer sentential descriptions of videos with a wide array of possible content.

Abstract: Despite a recent push towards large-scale object recognition, activity recognition remains limited to narrow domains and small vocabularies of actions. In this paper, we tackle the challenge of recognizing and describing activities "in-the-wild". We present a solution that takes a short video clip and outputs a brief sentence that sums up the main activity in the video, such as the actor, the action and its object. Unlike previous work, our approach works on out-of-domain actions: it does not require training videos of the exact activity. If it cannot find an accurate prediction for a pre-trained model, it finds a less specific answer that is also plausible from a pragmatic standpoint. We use semantic hierarchies learned from the data to help to choose an appropriate level of generalization, and priors learned from web-scale natural language corpora to penalize unlikely combinations of actors/actions/objects; we also use a web-scale language model to "fill in" novel verbs, i.e. when the verb does not appear in the training set. We evaluate our method on a large YouTube corpus and demonstrate it is able to generate short sentence descriptions of video clips better than baseline approaches.